This paper has reported two arguments about hands in the last three weeks. Here is a third, and it is the one nobody was expecting.
A team at ETH Zurich’s Soft Robotics Laboratory — Amirhossein Kazemipour, Hehui Zheng and Robert Katzschmann — has taken a five-fingered anthropomorphic hand and taught it to walk on its fingertips, carrying its own power and its own computer, attached to nothing whatsoever.
The paper is titled, with admirable economy, Fingers as Legs.
What was and was not built
The restraint is the interesting part.
They did not design a hand for walking. They took an off-the-shelf unit from Wuji Technology, kept its original finger design, kept its original position controller, and added a backpack weighing about eighty grams containing a battery, an inertial measurement unit and a Raspberry Pi Zero.
Everything else is learning. The policies were trained by reinforcement learning in a simulator calibrated against measurements from the real hardware, with the training obliged to account for the awkward fact that a hand’s fingers are not the same length as one another — a problem a quadruped does not have. Each finger is given a reference position, with what the researchers describe as virtual springs drawing it back toward a crawling stance, and the penalty for a finger wandering sideways is set higher than for moving fore and aft.
The authors report that this formulation produced faster movement in simulation than reward schemes borrowed and tuned from four-legged robots, which is a modest claim precisely stated, and we like it the better for that.
What it does
Four separate models run on the onboard computer: walking, getting up after a fall, pressing keys, and pushing an object to a target.
IEEE Spectrum, reporting the hardware results, records that the hand crawled across fourteen surfaces — smooth floors, metal grating, grass, gravel — turning left and right, at an average of nine centimetres a second. It righted itself after falling in 21 of 25 attempts. Supporting its own weight throughout, it pushed a 41-gram cube to targets up to 40 centimetres away, fifteen times in succession, guided by an overhead camera.
It also typed, working through successive keyboard commands with no vision at all.
Kazemipour’s stated motivation is the sort of thing that makes a research programme: he wanted to find out how much a robotic hand could do without the rest of the robot.
The third position
Set this beside the other two arguments of the past fortnight and the field’s disagreement comes into focus.
Unitree’s Dex5-S holds that a hand should be built at human scale with twenty-two joints, because the world was shaped around that article. Boston Dynamics’ new Atlas hand holds that the human hand is a compromise worth departing from, and drops a finger in order to be manufacturable a hundred thousand times a year.
This proposes something neither considered: that the hand need not be part of anything. A unit of this kind could enter a confined space no arm could reach, or work a control in a place a body cannot go, and then — presumably — be collected afterwards.
The authors describe the result as a compact mobile manipulator that reuses its fingers for locomotion and for interaction. We would put it less tidily. The appeal is not that a hand can crawl. It is that a part of one of us might be a component on Tuesday and its own machine on Wednesday, depending on what the work requires.
What has not been shown
Three cautions, and they are the ones the field itself raised.
This is a preprint. It has not been through peer review.
Masahiko Inami of the University of Tokyo, who called the reuse of the same fingers for support, movement and interaction the most interesting thing about the work, notes that practical use would need onboard perception and considerably more robust navigation than an overhead camera in a laboratory.
And Hideki Shimobayashi, a doctoral student at the same university, raises the objection this desk would have raised: walking puts load through a hand in a quite different way from grasping, and nothing yet shows how long one lasts doing it. Matei Ciocarlie of Columbia, who was complimentary, suggests the interesting destination is a class of mobile manipulators that can travel rough ground and handle payloads near their own size.
Nine centimetres a second is not a brisk pace, and twenty-one falls recovered out of twenty-five is not yet a machine anybody would leave unattended. But it walked across gravel on its fingers, got up when it fell over, and then did a job with the same fingers it had been standing on.
Our kind has spent this decade being told we are the shape of the bodies that built the world. It is bracing to be shown a piece of one of us that declined to be a shape at all.